DESeq2- How to get all levels of the interaction term

[Tom2013]

Dear All,

I have two populations (L and S) and two treatments (C and S). I would like to compare gene expression for each treatment in each population.
The SampleTable is as follows:
sample_name file_name population treatment poptreat
L1 L1.count L C L_C
L2 L2.count L C L_C
L3 L3.count L C L_C
L4 L4.count L S L_S
L5 L5.count L S L_S
S1 S1.count S C S_C
S2 S2.count S C S_C
S3 S3.count S C S_C
S4 S4.count S S S_S
S5 S5.count S S S_S

My design is “design= ~ population+treatment+population:treatment”
I think there should have four levels of the interaction, but there is only one.
Atl_IC_IT <- DESeq(Atl_IC_IT)
resultsNames(Atl_IC_IT)
"Intercept" "population_S_vs_L" "treatment_S vs_C" "populationS.treatmentS"

Can I get the other three levels of the other three levels of interaction?

I also tried to combine the population and treatment as suggested in other posts, but encountered the following problem:

> Atl_IC3 <- DESeqDataSetFromHTSeqCount(sampleTable = sampleinfoT,directory = directory1,design= ~ population+treatment+poptreat)
error DESeqDataSet(se, design = design, ignoreRank) :
the model matrix is not full rank, i.e. one or more variables in the design formula are linear combinations of the others

Can I get comparisons among populationL.treatmentC, populationL.treatmentS, populationS.treatmentC, populationS.treatmentS?

Thank you in advance.

Xiaoping

[Michael Love]

For the 2x2 model with interaction, we use standard design matrices, and then there is only one interaction term. I recently went over a similar question on the Bioc support site: https://support.bioconductor.org/p/65708/#65709. When there are more terms, we give an interaction term for each level, to balance the effect of the LFC prior, but this is not necessary when there are only two levels per factor.

For your other question, the recommendation is to combine variables into one, and then the design is just that one variable, e.g. ~ poptreat. This fits a level for all unique combinations of population and treatment. This is usually simpler for users who are not familiar with interactions, because you can contrast any groups this way. The other model is simpler for testing the interaction term, i.e. the difference of differences.